Implementation of a convolutional neural network used to identify wheezes and crackles in an audio file which is fed Mel-Spectrograms as inputs. During processing, audio clips are copied to 5 second long buffers, and are split into multiple segments if necessary, with zero padding added to fill the rest of the buffer. During training, Mel-Spectrograms are transposed and wrapped around the time-axis to help allow the network to learn to identify features occurring at arbitrary times within the recording. Data augmentation was employed in the form of audio stretching (speeding up and down) as well as Vocal Tract Length Perturbation, especially for the scarcer 'wheeze' and 'wheeze and crackles' classes. A one hot data labelling scheme was chosen as earlier attempts at using a multi-label scheme using a Sigmoid output layer resulted in poor training results (which in hindsight may have been caused by an excessively high learning rate). Currently, both the 'wheeze' and 'wheeze and crackles' classes pose the greatest challenge in the area of classification, and frequently produce false negatives, as indicated by the poor recall scores. Overall validation accuracy currently stands at roughly 70%.
